Anthropic has officially launched Opus 5, the newest version of its flagship Opus AI model. The release comes just two months after Opus 4.8.

Although Opus 5 is smaller than Anthropic’s recently released Fable 5 model, it may become the better choice for many users. 

The company says it is cheaper to run, faces fewer usage restrictions, and even outperforms Fable 5 on several benchmark tests.

The update also addresses one of the biggest complaints surrounding Fable 5 by easing some privacy-related limitations while introducing a smarter way to handle safety restrictions.

Opus 5

Anthropic has released new AI models at an impressive pace in 2026. Opus 4.8 became available on May 28. Then, in June, the company introduced Mythos 5, Fable 5, and Sonnet 5.

Now, Opus 5 joins the lineup, leaving Haiku as the only model still waiting for an upgrade to the 5-series family.

The quick rollout shows how aggressively Anthropic is refining its AI systems.

Instead of waiting many months between major releases, the company has adopted a much faster development cycle, bringing improvements to users more frequently.

Better Performance

One of the biggest surprises is that Opus 5 reportedly beats the larger Fable 5 model on several benchmarks shared by Anthropic.

Rather than focusing only on raw size, Anthropic says it concentrated on making Opus 5 better at checking its own work and improving its answers before presenting them to users.

According to the company, benchmark testing showed Opus 5 successfully created its own computer vision pipeline even when given an incomplete prompt. 

It is telling of stronger reasoning, better planning, and improved problem-solving compared to earlier versions. 

This focus on careful verification could make Opus 5 useful for developers, researchers, and professionals who rely on accurate AI-generated results.

Lower Costs

Anthropic says Opus 5 is less expensive than Fable 5 while still delivering stronger overall performance in several areas. 

That combination could make it an attractive option for companies looking to balance quality with operating costs.

For developers using Anthropic’s API, lower costs may also encourage larger-scale deployments without sacrificing capability. 

Privacy Improvements

Opus 5 doesn’t follow Anthropic’s 30-day data retention policy. Those earlier models store user data for up to 30 days, a concerning policy for organizations working with sensitive information.

By excluding Opus 5 from that retention requirement, Anthropic gives users another option when privacy is a higher priority.

For businesses handling confidential data, this difference could play a significant role when deciding which Anthropic model to use.

Safety Guardrails

While Opus 5 has fewer restrictions overall, Anthropic has not removed its safety systems. The company continues to limit requests involving potentially dangerous cybersecurity activities.

For example, Opus 5 will not help users scan compiled software binaries for security vulnerabilities. 

However, it can assist with reviewing source code for vulnerabilities because that task is generally associated with defensive security work, such as helping developers identify and fix weaknesses before attackers can exploit them.

This is a direct response to earlier criticism of Anthropic’s strict guardrails by cybersecurity experts.

Safety Interruptions

Anthropic also says Opus 5’s safety classifiers will activate far less often than those used with Fable 5.

According to the company, the safety system should intervene about 85% less frequently. 

That means users are less likely to encounter blocked requests during normal conversations and development work.

Instead of relying on broad restrictions, Anthropic appears to be moving toward more targeted safety measures that focus on genuinely risky requests while allowing everyday tasks to continue without unnecessary interruptions.

Automatic Fallbacks

Alongside Opus 5, Anthropic introduced a new beta feature called Automatic Fallbacks.

Previously, if a prompt triggered the company’s safety classifier, API users could receive an error instead of a response. 

With Automatic Fallbacks enabled, Anthropic will automatically route that request to a less capable model that can safely answer it.

As a result, users are more likely to receive a useful response instead of a failed request.

For developers building AI applications, this feature could improve reliability and create a smoother experience for end users.